/* * Copyright © 2016 Red Hat. * Copyright © 2016 Bas Nieuwenhuizen * * based in part on anv driver which is: * Copyright © 2015 Intel Corporation * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice (including the next * paragraph) shall be included in all copies or substantial portions of the * Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS * IN THE SOFTWARE. */ #include "radv_debug.h" #include "radv_private.h" #include "vk_format.h" #include "vk_util.h" #include "radv_radeon_winsys.h" #include "sid.h" #include "util/debug.h" #include "util/u_atomic.h" static unsigned radv_choose_tiling(struct radv_device *device, const struct radv_image_create_info *create_info) { const VkImageCreateInfo *pCreateInfo = create_info->vk_info; if (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR) { assert(pCreateInfo->samples <= 1); return RADEON_SURF_MODE_LINEAR_ALIGNED; } if (!vk_format_is_compressed(pCreateInfo->format) && !vk_format_is_depth_or_stencil(pCreateInfo->format) && device->physical_device->rad_info.chip_class <= GFX8) { /* this causes hangs in some VK CTS tests on GFX9. */ /* Textures with a very small height are recommended to be linear. */ if (pCreateInfo->imageType == VK_IMAGE_TYPE_1D || /* Only very thin and long 2D textures should benefit from * linear_aligned. */ (pCreateInfo->extent.width > 8 && pCreateInfo->extent.height <= 2)) return RADEON_SURF_MODE_LINEAR_ALIGNED; } /* MSAA resources must be 2D tiled. */ if (pCreateInfo->samples > 1) return RADEON_SURF_MODE_2D; return RADEON_SURF_MODE_2D; } static bool radv_use_tc_compat_htile_for_image(struct radv_device *device, const VkImageCreateInfo *pCreateInfo) { /* TC-compat HTILE is only available for GFX8+. */ if (device->physical_device->rad_info.chip_class < GFX8) return false; if ((pCreateInfo->usage & VK_IMAGE_USAGE_STORAGE_BIT) || (pCreateInfo->flags & VK_IMAGE_CREATE_EXTENDED_USAGE_BIT)) return false; if (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR) return false; if (pCreateInfo->mipLevels > 1) return false; /* FIXME: for some reason TC compat with 2/4/8 samples breaks some cts * tests - disable for now. On GFX10 D32_SFLOAT is affected as well. */ if (pCreateInfo->samples >= 2 && (pCreateInfo->format == VK_FORMAT_D32_SFLOAT_S8_UINT || (pCreateInfo->format == VK_FORMAT_D32_SFLOAT && device->physical_device->rad_info.chip_class == GFX10))) return false; /* GFX9 supports both 32-bit and 16-bit depth surfaces, while GFX8 only * supports 32-bit. Though, it's possible to enable TC-compat for * 16-bit depth surfaces if no Z planes are compressed. */ if (pCreateInfo->format != VK_FORMAT_D32_SFLOAT_S8_UINT && pCreateInfo->format != VK_FORMAT_D32_SFLOAT && pCreateInfo->format != VK_FORMAT_D16_UNORM) return false; if (pCreateInfo->flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) { const struct VkImageFormatListCreateInfoKHR *format_list = (const struct VkImageFormatListCreateInfoKHR *) vk_find_struct_const(pCreateInfo->pNext, IMAGE_FORMAT_LIST_CREATE_INFO_KHR); /* We have to ignore the existence of the list if viewFormatCount = 0 */ if (format_list && format_list->viewFormatCount) { /* compatibility is transitive, so we only need to check * one format with everything else. */ for (unsigned i = 0; i < format_list->viewFormatCount; ++i) { if (format_list->pViewFormats[i] == VK_FORMAT_UNDEFINED) continue; if (pCreateInfo->format != format_list->pViewFormats[i]) return false; } } else { return false; } } return true; } static bool radv_surface_has_scanout(struct radv_device *device, const struct radv_image_create_info *info) { if (info->scanout) return true; if (!info->bo_metadata) return false; if (device->physical_device->rad_info.chip_class >= GFX9) { return info->bo_metadata->u.gfx9.swizzle_mode == 0 || info->bo_metadata->u.gfx9.swizzle_mode % 4 == 2; } else { return info->bo_metadata->u.legacy.scanout; } } static bool radv_support_storage_dcc(const struct radv_physical_device *pdevice) { return pdevice->rad_info.chip_class >= GFX10 && (pdevice->instance->perftest_flags & RADV_PERFTEST_STORAGE_DCC); } static bool radv_use_dcc_for_image(struct radv_device *device, const struct radv_image *image, const struct radv_image_create_info *create_info, const VkImageCreateInfo *pCreateInfo) { bool dcc_compatible_formats; bool blendable; /* DCC (Delta Color Compression) is only available for GFX8+. */ if (device->physical_device->rad_info.chip_class < GFX8) return false; if (device->instance->debug_flags & RADV_DEBUG_NO_DCC) return false; if (image->shareable) return false; if (((pCreateInfo->usage & VK_IMAGE_USAGE_STORAGE_BIT) && !radv_support_storage_dcc(device->physical_device))|| (pCreateInfo->flags & VK_IMAGE_CREATE_EXTENDED_USAGE_BIT)) return false; if (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR) return false; if (vk_format_is_subsampled(pCreateInfo->format) || vk_format_get_plane_count(pCreateInfo->format) > 1) return false; /* TODO: Enable DCC for mipmaps on GFX9+. */ if ((pCreateInfo->arrayLayers > 1 || pCreateInfo->mipLevels > 1) && device->physical_device->rad_info.chip_class >= GFX9) return false; /* Do not enable DCC for mipmapped arrays because performance is worse. */ if (pCreateInfo->arrayLayers > 1 && pCreateInfo->mipLevels > 1) return false; if (radv_surface_has_scanout(device, create_info)) return false; /* FIXME: DCC for MSAA with 4x and 8x samples doesn't work yet, while * 2x can be enabled with an option. */ if (pCreateInfo->samples > 2 || (pCreateInfo->samples == 2 && !device->physical_device->dcc_msaa_allowed)) return false; /* Determine if the formats are DCC compatible. */ dcc_compatible_formats = radv_is_colorbuffer_format_supported(pCreateInfo->format, &blendable); if (pCreateInfo->flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) { const struct VkImageFormatListCreateInfoKHR *format_list = (const struct VkImageFormatListCreateInfoKHR *) vk_find_struct_const(pCreateInfo->pNext, IMAGE_FORMAT_LIST_CREATE_INFO_KHR); /* We have to ignore the existence of the list if viewFormatCount = 0 */ if (format_list && format_list->viewFormatCount) { /* compatibility is transitive, so we only need to check * one format with everything else. */ for (unsigned i = 0; i < format_list->viewFormatCount; ++i) { if (format_list->pViewFormats[i] == VK_FORMAT_UNDEFINED) continue; if (!radv_dcc_formats_compatible(pCreateInfo->format, format_list->pViewFormats[i])) dcc_compatible_formats = false; } } else { dcc_compatible_formats = false; } } if (!dcc_compatible_formats) return false; return true; } static bool radv_use_tc_compat_cmask_for_image(struct radv_device *device, struct radv_image *image) { if (!(device->instance->perftest_flags & RADV_PERFTEST_TC_COMPAT_CMASK)) return false; /* TC-compat CMASK is only available for GFX8+. */ if (device->physical_device->rad_info.chip_class < GFX8) return false; if (image->usage & VK_IMAGE_USAGE_STORAGE_BIT) return false; if (radv_image_has_dcc(image)) return false; if (!radv_image_has_cmask(image)) return false; return true; } static void radv_prefill_surface_from_metadata(struct radv_device *device, struct radeon_surf *surface, const struct radv_image_create_info *create_info) { const struct radeon_bo_metadata *md = create_info->bo_metadata; if (device->physical_device->rad_info.chip_class >= GFX9) { if (md->u.gfx9.swizzle_mode > 0) surface->flags |= RADEON_SURF_SET(RADEON_SURF_MODE_2D, MODE); else surface->flags |= RADEON_SURF_SET(RADEON_SURF_MODE_LINEAR_ALIGNED, MODE); surface->u.gfx9.surf.swizzle_mode = md->u.gfx9.swizzle_mode; } else { surface->u.legacy.pipe_config = md->u.legacy.pipe_config; surface->u.legacy.bankw = md->u.legacy.bankw; surface->u.legacy.bankh = md->u.legacy.bankh; surface->u.legacy.tile_split = md->u.legacy.tile_split; surface->u.legacy.mtilea = md->u.legacy.mtilea; surface->u.legacy.num_banks = md->u.legacy.num_banks; if (md->u.legacy.macrotile == RADEON_LAYOUT_TILED) surface->flags |= RADEON_SURF_SET(RADEON_SURF_MODE_2D, MODE); else if (md->u.legacy.microtile == RADEON_LAYOUT_TILED) surface->flags |= RADEON_SURF_SET(RADEON_SURF_MODE_1D, MODE); else surface->flags |= RADEON_SURF_SET(RADEON_SURF_MODE_LINEAR_ALIGNED, MODE); } } static int radv_init_surface(struct radv_device *device, const struct radv_image *image, struct radeon_surf *surface, unsigned plane_id, const struct radv_image_create_info *create_info) { const VkImageCreateInfo *pCreateInfo = create_info->vk_info; unsigned array_mode = radv_choose_tiling(device, create_info); VkFormat format = vk_format_get_plane_format(pCreateInfo->format, plane_id); const struct vk_format_description *desc = vk_format_description(format); bool is_depth, is_stencil; is_depth = vk_format_has_depth(desc); is_stencil = vk_format_has_stencil(desc); surface->blk_w = vk_format_get_blockwidth(format); surface->blk_h = vk_format_get_blockheight(format); surface->bpe = vk_format_get_blocksize(vk_format_depth_only(format)); /* align byte per element on dword */ if (surface->bpe == 3) { surface->bpe = 4; } if (create_info->bo_metadata) { radv_prefill_surface_from_metadata(device, surface, create_info); } else { surface->flags = RADEON_SURF_SET(array_mode, MODE); } switch (pCreateInfo->imageType){ case VK_IMAGE_TYPE_1D: if (pCreateInfo->arrayLayers > 1) surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_1D_ARRAY, TYPE); else surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_1D, TYPE); break; case VK_IMAGE_TYPE_2D: if (pCreateInfo->arrayLayers > 1) surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_2D_ARRAY, TYPE); else surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_2D, TYPE); break; case VK_IMAGE_TYPE_3D: surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_3D, TYPE); break; default: unreachable("unhandled image type"); } if (is_depth) { surface->flags |= RADEON_SURF_ZBUFFER; if (radv_use_tc_compat_htile_for_image(device, pCreateInfo)) surface->flags |= RADEON_SURF_TC_COMPATIBLE_HTILE; } if (is_stencil) surface->flags |= RADEON_SURF_SBUFFER; if (device->physical_device->rad_info.chip_class >= GFX9 && pCreateInfo->imageType == VK_IMAGE_TYPE_3D && vk_format_get_blocksizebits(pCreateInfo->format) == 128 && vk_format_is_compressed(pCreateInfo->format)) surface->flags |= RADEON_SURF_NO_RENDER_TARGET; surface->flags |= RADEON_SURF_OPTIMIZE_FOR_SPACE; if (!radv_use_dcc_for_image(device, image, create_info, pCreateInfo)) surface->flags |= RADEON_SURF_DISABLE_DCC; if (radv_surface_has_scanout(device, create_info)) surface->flags |= RADEON_SURF_SCANOUT; return 0; } static uint32_t si_get_bo_metadata_word1(struct radv_device *device) { return (ATI_VENDOR_ID << 16) | device->physical_device->rad_info.pci_id; } static inline unsigned si_tile_mode_index(const struct radv_image_plane *plane, unsigned level, bool stencil) { if (stencil) return plane->surface.u.legacy.stencil_tiling_index[level]; else return plane->surface.u.legacy.tiling_index[level]; } static unsigned radv_map_swizzle(unsigned swizzle) { switch (swizzle) { case VK_SWIZZLE_Y: return V_008F0C_SQ_SEL_Y; case VK_SWIZZLE_Z: return V_008F0C_SQ_SEL_Z; case VK_SWIZZLE_W: return V_008F0C_SQ_SEL_W; case VK_SWIZZLE_0: return V_008F0C_SQ_SEL_0; case VK_SWIZZLE_1: return V_008F0C_SQ_SEL_1; default: /* VK_SWIZZLE_X */ return V_008F0C_SQ_SEL_X; } } static void radv_make_buffer_descriptor(struct radv_device *device, struct radv_buffer *buffer, VkFormat vk_format, unsigned offset, unsigned range, uint32_t *state) { const struct vk_format_description *desc; unsigned stride; uint64_t gpu_address = radv_buffer_get_va(buffer->bo); uint64_t va = gpu_address + buffer->offset; unsigned num_format, data_format; int first_non_void; desc = vk_format_description(vk_format); first_non_void = vk_format_get_first_non_void_channel(vk_format); stride = desc->block.bits / 8; va += offset; state[0] = va; state[1] = S_008F04_BASE_ADDRESS_HI(va >> 32) | S_008F04_STRIDE(stride); if (device->physical_device->rad_info.chip_class != GFX8 && stride) { range /= stride; } state[2] = range; state[3] = S_008F0C_DST_SEL_X(radv_map_swizzle(desc->swizzle[0])) | S_008F0C_DST_SEL_Y(radv_map_swizzle(desc->swizzle[1])) | S_008F0C_DST_SEL_Z(radv_map_swizzle(desc->swizzle[2])) | S_008F0C_DST_SEL_W(radv_map_swizzle(desc->swizzle[3])); if (device->physical_device->rad_info.chip_class >= GFX10) { const struct gfx10_format *fmt = &gfx10_format_table[vk_format]; /* OOB_SELECT chooses the out-of-bounds check: * - 0: (index >= NUM_RECORDS) || (offset >= STRIDE) * - 1: index >= NUM_RECORDS * - 2: NUM_RECORDS == 0 * - 3: if SWIZZLE_ENABLE == 0: offset >= NUM_RECORDS * else: swizzle_address >= NUM_RECORDS */ state[3] |= S_008F0C_FORMAT(fmt->img_format) | S_008F0C_OOB_SELECT(0) | S_008F0C_RESOURCE_LEVEL(1); } else { num_format = radv_translate_buffer_numformat(desc, first_non_void); data_format = radv_translate_buffer_dataformat(desc, first_non_void); assert(data_format != V_008F0C_BUF_DATA_FORMAT_INVALID); assert(num_format != ~0); state[3] |= S_008F0C_NUM_FORMAT(num_format) | S_008F0C_DATA_FORMAT(data_format); } } static void si_set_mutable_tex_desc_fields(struct radv_device *device, struct radv_image *image, const struct legacy_surf_level *base_level_info, unsigned plane_id, unsigned base_level, unsigned first_level, unsigned block_width, bool is_stencil, bool is_storage_image, bool disable_compression, uint32_t *state) { struct radv_image_plane *plane = &image->planes[plane_id]; uint64_t gpu_address = image->bo ? radv_buffer_get_va(image->bo) + image->offset : 0; uint64_t va = gpu_address + plane->offset; enum chip_class chip_class = device->physical_device->rad_info.chip_class; uint64_t meta_va = 0; if (chip_class >= GFX9) { if (is_stencil) va += plane->surface.u.gfx9.stencil_offset; else va += plane->surface.u.gfx9.surf_offset; } else va += base_level_info->offset; state[0] = va >> 8; if (chip_class >= GFX9 || base_level_info->mode == RADEON_SURF_MODE_2D) state[0] |= plane->surface.tile_swizzle; state[1] &= C_008F14_BASE_ADDRESS_HI; state[1] |= S_008F14_BASE_ADDRESS_HI(va >> 40); if (chip_class >= GFX8) { state[6] &= C_008F28_COMPRESSION_EN; state[7] = 0; if (!disable_compression && radv_dcc_enabled(image, first_level)) { meta_va = gpu_address + image->dcc_offset; if (chip_class <= GFX8) meta_va += base_level_info->dcc_offset; unsigned dcc_tile_swizzle = plane->surface.tile_swizzle << 8; dcc_tile_swizzle &= plane->surface.dcc_alignment - 1; meta_va |= dcc_tile_swizzle; } else if (!disable_compression && radv_image_is_tc_compat_htile(image)) { meta_va = gpu_address + image->htile_offset; } if (meta_va) { state[6] |= S_008F28_COMPRESSION_EN(1); if (chip_class <= GFX9) state[7] = meta_va >> 8; } } if (chip_class >= GFX10) { state[3] &= C_00A00C_SW_MODE; if (is_stencil) { state[3] |= S_00A00C_SW_MODE(plane->surface.u.gfx9.stencil.swizzle_mode); } else { state[3] |= S_00A00C_SW_MODE(plane->surface.u.gfx9.surf.swizzle_mode); } state[6] &= C_00A018_META_DATA_ADDRESS_LO & C_00A018_META_PIPE_ALIGNED; if (meta_va) { struct gfx9_surf_meta_flags meta; if (image->dcc_offset) meta = plane->surface.u.gfx9.dcc; else meta = plane->surface.u.gfx9.htile; state[6] |= S_00A018_META_PIPE_ALIGNED(meta.pipe_aligned) | S_00A018_META_DATA_ADDRESS_LO(meta_va >> 8); } state[7] = meta_va >> 16; } else if (chip_class == GFX9) { state[3] &= C_008F1C_SW_MODE; state[4] &= C_008F20_PITCH; if (is_stencil) { state[3] |= S_008F1C_SW_MODE(plane->surface.u.gfx9.stencil.swizzle_mode); state[4] |= S_008F20_PITCH(plane->surface.u.gfx9.stencil.epitch); } else { state[3] |= S_008F1C_SW_MODE(plane->surface.u.gfx9.surf.swizzle_mode); state[4] |= S_008F20_PITCH(plane->surface.u.gfx9.surf.epitch); } state[5] &= C_008F24_META_DATA_ADDRESS & C_008F24_META_PIPE_ALIGNED & C_008F24_META_RB_ALIGNED; if (meta_va) { struct gfx9_surf_meta_flags meta; if (image->dcc_offset) meta = plane->surface.u.gfx9.dcc; else meta = plane->surface.u.gfx9.htile; state[5] |= S_008F24_META_DATA_ADDRESS(meta_va >> 40) | S_008F24_META_PIPE_ALIGNED(meta.pipe_aligned) | S_008F24_META_RB_ALIGNED(meta.rb_aligned); } } else { /* GFX6-GFX8 */ unsigned pitch = base_level_info->nblk_x * block_width; unsigned index = si_tile_mode_index(plane, base_level, is_stencil); state[3] &= C_008F1C_TILING_INDEX; state[3] |= S_008F1C_TILING_INDEX(index); state[4] &= C_008F20_PITCH; state[4] |= S_008F20_PITCH(pitch - 1); } } static unsigned radv_tex_dim(VkImageType image_type, VkImageViewType view_type, unsigned nr_layers, unsigned nr_samples, bool is_storage_image, bool gfx9) { if (view_type == VK_IMAGE_VIEW_TYPE_CUBE || view_type == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY) return is_storage_image ? V_008F1C_SQ_RSRC_IMG_2D_ARRAY : V_008F1C_SQ_RSRC_IMG_CUBE; /* GFX9 allocates 1D textures as 2D. */ if (gfx9 && image_type == VK_IMAGE_TYPE_1D) image_type = VK_IMAGE_TYPE_2D; switch (image_type) { case VK_IMAGE_TYPE_1D: return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_1D_ARRAY : V_008F1C_SQ_RSRC_IMG_1D; case VK_IMAGE_TYPE_2D: if (nr_samples > 1) return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY : V_008F1C_SQ_RSRC_IMG_2D_MSAA; else return nr_layers > 1 ? V_008F1C_SQ_RSRC_IMG_2D_ARRAY : V_008F1C_SQ_RSRC_IMG_2D; case VK_IMAGE_TYPE_3D: if (view_type == VK_IMAGE_VIEW_TYPE_3D) return V_008F1C_SQ_RSRC_IMG_3D; else return V_008F1C_SQ_RSRC_IMG_2D_ARRAY; default: unreachable("illegal image type"); } } static unsigned gfx9_border_color_swizzle(const enum vk_swizzle swizzle[4]) { unsigned bc_swizzle = V_008F20_BC_SWIZZLE_XYZW; if (swizzle[3] == VK_SWIZZLE_X) { /* For the pre-defined border color values (white, opaque * black, transparent black), the only thing that matters is * that the alpha channel winds up in the correct place * (because the RGB channels are all the same) so either of * these enumerations will work. */ if (swizzle[2] == VK_SWIZZLE_Y) bc_swizzle = V_008F20_BC_SWIZZLE_WZYX; else bc_swizzle = V_008F20_BC_SWIZZLE_WXYZ; } else if (swizzle[0] == VK_SWIZZLE_X) { if (swizzle[1] == VK_SWIZZLE_Y) bc_swizzle = V_008F20_BC_SWIZZLE_XYZW; else bc_swizzle = V_008F20_BC_SWIZZLE_XWYZ; } else if (swizzle[1] == VK_SWIZZLE_X) { bc_swizzle = V_008F20_BC_SWIZZLE_YXWZ; } else if (swizzle[2] == VK_SWIZZLE_X) { bc_swizzle = V_008F20_BC_SWIZZLE_ZYXW; } return bc_swizzle; } static bool vi_alpha_is_on_msb(struct radv_device *device, VkFormat format) { const struct vk_format_description *desc = vk_format_description(format); if (device->physical_device->rad_info.chip_class >= GFX10 && desc->nr_channels == 1) return desc->swizzle[3] == VK_SWIZZLE_X; return radv_translate_colorswap(format, false) <= 1; } /** * Build the sampler view descriptor for a texture (GFX10). */ static void gfx10_make_texture_descriptor(struct radv_device *device, struct radv_image *image, bool is_storage_image, VkImageViewType view_type, VkFormat vk_format, const VkComponentMapping *mapping, unsigned first_level, unsigned last_level, unsigned first_layer, unsigned last_layer, unsigned width, unsigned height, unsigned depth, uint32_t *state, uint32_t *fmask_state) { const struct vk_format_description *desc; enum vk_swizzle swizzle[4]; unsigned img_format; unsigned type; desc = vk_format_description(vk_format); img_format = gfx10_format_table[vk_format].img_format; if (desc->colorspace == VK_FORMAT_COLORSPACE_ZS) { const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0}; vk_format_compose_swizzles(mapping, swizzle_xxxx, swizzle); } else { vk_format_compose_swizzles(mapping, desc->swizzle, swizzle); } type = radv_tex_dim(image->type, view_type, image->info.array_size, image->info.samples, is_storage_image, device->physical_device->rad_info.chip_class == GFX9); if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) { height = 1; depth = image->info.array_size; } else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY || type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) { if (view_type != VK_IMAGE_VIEW_TYPE_3D) depth = image->info.array_size; } else if (type == V_008F1C_SQ_RSRC_IMG_CUBE) depth = image->info.array_size / 6; state[0] = 0; state[1] = S_00A004_FORMAT(img_format) | S_00A004_WIDTH_LO(width - 1); state[2] = S_00A008_WIDTH_HI((width - 1) >> 2) | S_00A008_HEIGHT(height - 1) | S_00A008_RESOURCE_LEVEL(1); state[3] = S_00A00C_DST_SEL_X(radv_map_swizzle(swizzle[0])) | S_00A00C_DST_SEL_Y(radv_map_swizzle(swizzle[1])) | S_00A00C_DST_SEL_Z(radv_map_swizzle(swizzle[2])) | S_00A00C_DST_SEL_W(radv_map_swizzle(swizzle[3])) | S_00A00C_BASE_LEVEL(image->info.samples > 1 ? 0 : first_level) | S_00A00C_LAST_LEVEL(image->info.samples > 1 ? util_logbase2(image->info.samples) : last_level) | S_00A00C_BC_SWIZZLE(gfx9_border_color_swizzle(swizzle)) | S_00A00C_TYPE(type); /* Depth is the the last accessible layer on gfx9+. The hw doesn't need * to know the total number of layers. */ state[4] = S_00A010_DEPTH(type == V_008F1C_SQ_RSRC_IMG_3D ? depth - 1 : last_layer) | S_00A010_BASE_ARRAY(first_layer); state[5] = S_00A014_ARRAY_PITCH(!!(type == V_008F1C_SQ_RSRC_IMG_3D)) | S_00A014_MAX_MIP(image->info.samples > 1 ? util_logbase2(image->info.samples) : image->info.levels - 1) | S_00A014_PERF_MOD(4); state[6] = 0; state[7] = 0; if (radv_dcc_enabled(image, first_level)) { state[6] |= S_00A018_MAX_UNCOMPRESSED_BLOCK_SIZE(V_028C78_MAX_BLOCK_SIZE_256B) | S_00A018_MAX_COMPRESSED_BLOCK_SIZE(V_028C78_MAX_BLOCK_SIZE_128B) | S_00A018_ALPHA_IS_ON_MSB(vi_alpha_is_on_msb(device, vk_format)); } /* Initialize the sampler view for FMASK. */ if (radv_image_has_fmask(image)) { uint64_t gpu_address = radv_buffer_get_va(image->bo); uint32_t format; uint64_t va; assert(image->plane_count == 1); va = gpu_address + image->offset + image->fmask_offset; switch (image->info.samples) { case 2: format = V_008F0C_IMG_FORMAT_FMASK8_S2_F2; break; case 4: format = V_008F0C_IMG_FORMAT_FMASK8_S4_F4; break; case 8: format = V_008F0C_IMG_FORMAT_FMASK32_S8_F8; break; default: unreachable("invalid nr_samples"); } fmask_state[0] = (va >> 8) | image->planes[0].surface.fmask_tile_swizzle; fmask_state[1] = S_00A004_BASE_ADDRESS_HI(va >> 40) | S_00A004_FORMAT(format) | S_00A004_WIDTH_LO(width - 1); fmask_state[2] = S_00A008_WIDTH_HI((width - 1) >> 2) | S_00A008_HEIGHT(height - 1) | S_00A008_RESOURCE_LEVEL(1); fmask_state[3] = S_00A00C_DST_SEL_X(V_008F1C_SQ_SEL_X) | S_00A00C_DST_SEL_Y(V_008F1C_SQ_SEL_X) | S_00A00C_DST_SEL_Z(V_008F1C_SQ_SEL_X) | S_00A00C_DST_SEL_W(V_008F1C_SQ_SEL_X) | S_00A00C_SW_MODE(image->planes[0].surface.u.gfx9.fmask.swizzle_mode) | S_00A00C_TYPE(radv_tex_dim(image->type, view_type, image->info.array_size, 0, false, false)); fmask_state[4] = S_00A010_DEPTH(last_layer) | S_00A010_BASE_ARRAY(first_layer); fmask_state[5] = 0; fmask_state[6] = S_00A018_META_PIPE_ALIGNED(image->planes[0].surface.u.gfx9.cmask.pipe_aligned); fmask_state[7] = 0; } else if (fmask_state) memset(fmask_state, 0, 8 * 4); } /** * Build the sampler view descriptor for a texture (SI-GFX9) */ static void si_make_texture_descriptor(struct radv_device *device, struct radv_image *image, bool is_storage_image, VkImageViewType view_type, VkFormat vk_format, const VkComponentMapping *mapping, unsigned first_level, unsigned last_level, unsigned first_layer, unsigned last_layer, unsigned width, unsigned height, unsigned depth, uint32_t *state, uint32_t *fmask_state) { const struct vk_format_description *desc; enum vk_swizzle swizzle[4]; int first_non_void; unsigned num_format, data_format, type; desc = vk_format_description(vk_format); if (desc->colorspace == VK_FORMAT_COLORSPACE_ZS) { const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0}; vk_format_compose_swizzles(mapping, swizzle_xxxx, swizzle); } else { vk_format_compose_swizzles(mapping, desc->swizzle, swizzle); } first_non_void = vk_format_get_first_non_void_channel(vk_format); num_format = radv_translate_tex_numformat(vk_format, desc, first_non_void); if (num_format == ~0) { num_format = 0; } data_format = radv_translate_tex_dataformat(vk_format, desc, first_non_void); if (data_format == ~0) { data_format = 0; } /* S8 with either Z16 or Z32 HTILE need a special format. */ if (device->physical_device->rad_info.chip_class == GFX9 && vk_format == VK_FORMAT_S8_UINT && radv_image_is_tc_compat_htile(image)) { if (image->vk_format == VK_FORMAT_D32_SFLOAT_S8_UINT) data_format = V_008F14_IMG_DATA_FORMAT_S8_32; else if (image->vk_format == VK_FORMAT_D16_UNORM_S8_UINT) data_format = V_008F14_IMG_DATA_FORMAT_S8_16; } type = radv_tex_dim(image->type, view_type, image->info.array_size, image->info.samples, is_storage_image, device->physical_device->rad_info.chip_class == GFX9); if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) { height = 1; depth = image->info.array_size; } else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY || type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) { if (view_type != VK_IMAGE_VIEW_TYPE_3D) depth = image->info.array_size; } else if (type == V_008F1C_SQ_RSRC_IMG_CUBE) depth = image->info.array_size / 6; state[0] = 0; state[1] = (S_008F14_DATA_FORMAT(data_format) | S_008F14_NUM_FORMAT(num_format)); state[2] = (S_008F18_WIDTH(width - 1) | S_008F18_HEIGHT(height - 1) | S_008F18_PERF_MOD(4)); state[3] = (S_008F1C_DST_SEL_X(radv_map_swizzle(swizzle[0])) | S_008F1C_DST_SEL_Y(radv_map_swizzle(swizzle[1])) | S_008F1C_DST_SEL_Z(radv_map_swizzle(swizzle[2])) | S_008F1C_DST_SEL_W(radv_map_swizzle(swizzle[3])) | S_008F1C_BASE_LEVEL(image->info.samples > 1 ? 0 : first_level) | S_008F1C_LAST_LEVEL(image->info.samples > 1 ? util_logbase2(image->info.samples) : last_level) | S_008F1C_TYPE(type)); state[4] = 0; state[5] = S_008F24_BASE_ARRAY(first_layer); state[6] = 0; state[7] = 0; if (device->physical_device->rad_info.chip_class == GFX9) { unsigned bc_swizzle = gfx9_border_color_swizzle(swizzle); /* Depth is the last accessible layer on Gfx9. * The hw doesn't need to know the total number of layers. */ if (type == V_008F1C_SQ_RSRC_IMG_3D) state[4] |= S_008F20_DEPTH(depth - 1); else state[4] |= S_008F20_DEPTH(last_layer); state[4] |= S_008F20_BC_SWIZZLE(bc_swizzle); state[5] |= S_008F24_MAX_MIP(image->info.samples > 1 ? util_logbase2(image->info.samples) : image->info.levels - 1); } else { state[3] |= S_008F1C_POW2_PAD(image->info.levels > 1); state[4] |= S_008F20_DEPTH(depth - 1); state[5] |= S_008F24_LAST_ARRAY(last_layer); } if (image->dcc_offset) { state[6] = S_008F28_ALPHA_IS_ON_MSB(vi_alpha_is_on_msb(device, vk_format)); } else { /* The last dword is unused by hw. The shader uses it to clear * bits in the first dword of sampler state. */ if (device->physical_device->rad_info.chip_class <= GFX7 && image->info.samples <= 1) { if (first_level == last_level) state[7] = C_008F30_MAX_ANISO_RATIO; else state[7] = 0xffffffff; } } /* Initialize the sampler view for FMASK. */ if (radv_image_has_fmask(image)) { uint32_t fmask_format, num_format; uint64_t gpu_address = radv_buffer_get_va(image->bo); uint64_t va; assert(image->plane_count == 1); va = gpu_address + image->offset + image->fmask_offset; if (device->physical_device->rad_info.chip_class == GFX9) { fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK; switch (image->info.samples) { case 2: num_format = V_008F14_IMG_FMASK_8_2_2; break; case 4: num_format = V_008F14_IMG_FMASK_8_4_4; break; case 8: num_format = V_008F14_IMG_FMASK_32_8_8; break; default: unreachable("invalid nr_samples"); } } else { switch (image->info.samples) { case 2: fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2; break; case 4: fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4; break; case 8: fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8; break; default: assert(0); fmask_format = V_008F14_IMG_DATA_FORMAT_INVALID; } num_format = V_008F14_IMG_NUM_FORMAT_UINT; } fmask_state[0] = va >> 8; fmask_state[0] |= image->planes[0].surface.fmask_tile_swizzle; fmask_state[1] = S_008F14_BASE_ADDRESS_HI(va >> 40) | S_008F14_DATA_FORMAT(fmask_format) | S_008F14_NUM_FORMAT(num_format); fmask_state[2] = S_008F18_WIDTH(width - 1) | S_008F18_HEIGHT(height - 1); fmask_state[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X) | S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X) | S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X) | S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X) | S_008F1C_TYPE(radv_tex_dim(image->type, view_type, image->info.array_size, 0, false, false)); fmask_state[4] = 0; fmask_state[5] = S_008F24_BASE_ARRAY(first_layer); fmask_state[6] = 0; fmask_state[7] = 0; if (device->physical_device->rad_info.chip_class == GFX9) { fmask_state[3] |= S_008F1C_SW_MODE(image->planes[0].surface.u.gfx9.fmask.swizzle_mode); fmask_state[4] |= S_008F20_DEPTH(last_layer) | S_008F20_PITCH(image->planes[0].surface.u.gfx9.fmask.epitch); fmask_state[5] |= S_008F24_META_PIPE_ALIGNED(image->planes[0].surface.u.gfx9.cmask.pipe_aligned) | S_008F24_META_RB_ALIGNED(image->planes[0].surface.u.gfx9.cmask.rb_aligned); if (radv_image_is_tc_compat_cmask(image)) { va = gpu_address + image->offset + image->cmask_offset; fmask_state[5] |= S_008F24_META_DATA_ADDRESS(va >> 40); fmask_state[6] |= S_008F28_COMPRESSION_EN(1); fmask_state[7] |= va >> 8; } } else { fmask_state[3] |= S_008F1C_TILING_INDEX(image->planes[0].surface.u.legacy.fmask.tiling_index); fmask_state[4] |= S_008F20_DEPTH(depth - 1) | S_008F20_PITCH(image->planes[0].surface.u.legacy.fmask.pitch_in_pixels - 1); fmask_state[5] |= S_008F24_LAST_ARRAY(last_layer); if (radv_image_is_tc_compat_cmask(image)) { va = gpu_address + image->offset + image->cmask_offset; fmask_state[6] |= S_008F28_COMPRESSION_EN(1); fmask_state[7] |= va >> 8; } } } else if (fmask_state) memset(fmask_state, 0, 8 * 4); } static void radv_make_texture_descriptor(struct radv_device *device, struct radv_image *image, bool is_storage_image, VkImageViewType view_type, VkFormat vk_format, const VkComponentMapping *mapping, unsigned first_level, unsigned last_level, unsigned first_layer, unsigned last_layer, unsigned width, unsigned height, unsigned depth, uint32_t *state, uint32_t *fmask_state) { if (device->physical_device->rad_info.chip_class >= GFX10) { gfx10_make_texture_descriptor(device, image, is_storage_image, view_type, vk_format, mapping, first_level, last_level, first_layer, last_layer, width, height, depth, state, fmask_state); } else { si_make_texture_descriptor(device, image, is_storage_image, view_type, vk_format, mapping, first_level, last_level, first_layer, last_layer, width, height, depth, state, fmask_state); } } static void radv_query_opaque_metadata(struct radv_device *device, struct radv_image *image, struct radeon_bo_metadata *md) { static const VkComponentMapping fixedmapping; uint32_t desc[8], i; assert(image->plane_count == 1); /* Metadata image format format version 1: * [0] = 1 (metadata format identifier) * [1] = (VENDOR_ID << 16) | PCI_ID * [2:9] = image descriptor for the whole resource * [2] is always 0, because the base address is cleared * [9] is the DCC offset bits [39:8] from the beginning of * the buffer * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level */ md->metadata[0] = 1; /* metadata image format version 1 */ /* TILE_MODE_INDEX is ambiguous without a PCI ID. */ md->metadata[1] = si_get_bo_metadata_word1(device); radv_make_texture_descriptor(device, image, false, (VkImageViewType)image->type, image->vk_format, &fixedmapping, 0, image->info.levels - 1, 0, image->info.array_size - 1, image->info.width, image->info.height, image->info.depth, desc, NULL); si_set_mutable_tex_desc_fields(device, image, &image->planes[0].surface.u.legacy.level[0], 0, 0, 0, image->planes[0].surface.blk_w, false, false, false, desc); /* Clear the base address and set the relative DCC offset. */ desc[0] = 0; desc[1] &= C_008F14_BASE_ADDRESS_HI; desc[7] = image->dcc_offset >> 8; /* Dwords [2:9] contain the image descriptor. */ memcpy(&md->metadata[2], desc, sizeof(desc)); /* Dwords [10:..] contain the mipmap level offsets. */ if (device->physical_device->rad_info.chip_class <= GFX8) { for (i = 0; i <= image->info.levels - 1; i++) md->metadata[10+i] = image->planes[0].surface.u.legacy.level[i].offset >> 8; md->size_metadata = (11 + image->info.levels - 1) * 4; } else md->size_metadata = 10 * 4; } void radv_init_metadata(struct radv_device *device, struct radv_image *image, struct radeon_bo_metadata *metadata) { struct radeon_surf *surface = &image->planes[0].surface; memset(metadata, 0, sizeof(*metadata)); if (device->physical_device->rad_info.chip_class >= GFX9) { metadata->u.gfx9.swizzle_mode = surface->u.gfx9.surf.swizzle_mode; } else { metadata->u.legacy.microtile = surface->u.legacy.level[0].mode >= RADEON_SURF_MODE_1D ? RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR; metadata->u.legacy.macrotile = surface->u.legacy.level[0].mode >= RADEON_SURF_MODE_2D ? RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR; metadata->u.legacy.pipe_config = surface->u.legacy.pipe_config; metadata->u.legacy.bankw = surface->u.legacy.bankw; metadata->u.legacy.bankh = surface->u.legacy.bankh; metadata->u.legacy.tile_split = surface->u.legacy.tile_split; metadata->u.legacy.mtilea = surface->u.legacy.mtilea; metadata->u.legacy.num_banks = surface->u.legacy.num_banks; metadata->u.legacy.stride = surface->u.legacy.level[0].nblk_x * surface->bpe; metadata->u.legacy.scanout = (surface->flags & RADEON_SURF_SCANOUT) != 0; } radv_query_opaque_metadata(device, image, metadata); } void radv_image_override_offset_stride(struct radv_device *device, struct radv_image *image, uint64_t offset, uint32_t stride) { struct radeon_surf *surface = &image->planes[0].surface; unsigned bpe = vk_format_get_blocksizebits(image->vk_format) / 8; if (device->physical_device->rad_info.chip_class >= GFX9) { if (stride) { surface->u.gfx9.surf_pitch = stride; surface->u.gfx9.surf_slice_size = (uint64_t)stride * surface->u.gfx9.surf_height * bpe; } surface->u.gfx9.surf_offset = offset; } else { surface->u.legacy.level[0].nblk_x = stride; surface->u.legacy.level[0].slice_size_dw = ((uint64_t)stride * surface->u.legacy.level[0].nblk_y * bpe) / 4; if (offset) { for (unsigned i = 0; i < ARRAY_SIZE(surface->u.legacy.level); ++i) surface->u.legacy.level[i].offset += offset; } } } static void radv_image_alloc_fmask(struct radv_device *device, struct radv_image *image) { unsigned fmask_alignment = image->planes[0].surface.fmask_alignment; image->fmask_offset = align64(image->size, fmask_alignment); image->size = image->fmask_offset + image->planes[0].surface.fmask_size; image->alignment = MAX2(image->alignment, fmask_alignment); } static void radv_image_alloc_cmask(struct radv_device *device, struct radv_image *image) { unsigned cmask_alignment = image->planes[0].surface.cmask_alignment; unsigned cmask_size = image->planes[0].surface.cmask_size; uint32_t clear_value_size = 0; if (!cmask_size) return; assert(cmask_alignment); image->cmask_offset = align64(image->size, cmask_alignment); /* + 8 for storing the clear values */ if (!image->clear_value_offset) { image->clear_value_offset = image->cmask_offset + cmask_size; clear_value_size = 8; } image->size = image->cmask_offset + cmask_size + clear_value_size; image->alignment = MAX2(image->alignment, cmask_alignment); } static void radv_image_alloc_dcc(struct radv_image *image) { assert(image->plane_count == 1); image->dcc_offset = align64(image->size, image->planes[0].surface.dcc_alignment); /* + 24 for storing the clear values + fce pred + dcc pred for each mip */ image->clear_value_offset = image->dcc_offset + image->planes[0].surface.dcc_size; image->fce_pred_offset = image->clear_value_offset + 8 * image->info.levels; image->dcc_pred_offset = image->clear_value_offset + 16 * image->info.levels; image->size = image->dcc_offset + image->planes[0].surface.dcc_size + 24 * image->info.levels; image->alignment = MAX2(image->alignment, image->planes[0].surface.dcc_alignment); } static void radv_image_alloc_htile(struct radv_device *device, struct radv_image *image) { image->htile_offset = align64(image->size, image->planes[0].surface.htile_alignment); /* + 8 for storing the clear values */ image->clear_value_offset = image->htile_offset + image->planes[0].surface.htile_size; image->size = image->clear_value_offset + 8; if (radv_image_is_tc_compat_htile(image) && device->physical_device->has_tc_compat_zrange_bug) { /* Metadata for the TC-compatible HTILE hardware bug which * have to be fixed by updating ZRANGE_PRECISION when doing * fast depth clears to 0.0f. */ image->tc_compat_zrange_offset = image->size; image->size = image->tc_compat_zrange_offset + 4; } image->alignment = align64(image->alignment, image->planes[0].surface.htile_alignment); } static inline bool radv_image_can_enable_dcc_or_cmask(struct radv_image *image) { if (image->info.samples <= 1 && image->info.width * image->info.height <= 512 * 512) { /* Do not enable CMASK or DCC for small surfaces where the cost * of the eliminate pass can be higher than the benefit of fast * clear. RadeonSI does this, but the image threshold is * different. */ return false; } return image->usage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT && (image->exclusive || image->queue_family_mask == 1); } static inline bool radv_image_can_enable_dcc(struct radv_device *device, struct radv_image *image) { if (!radv_image_can_enable_dcc_or_cmask(image) || !radv_image_has_dcc(image)) return false; /* On GFX8, DCC layers can be interleaved and it's currently only * enabled if slice size is equal to the per slice fast clear size * because the driver assumes that portions of multiple layers are * contiguous during fast clears. */ if (image->info.array_size > 1) { const struct legacy_surf_level *surf_level = &image->planes[0].surface.u.legacy.level[0]; assert(device->physical_device->rad_info.chip_class == GFX8); if (image->planes[0].surface.dcc_slice_size != surf_level->dcc_fast_clear_size) return false; } return true; } static inline bool radv_image_can_enable_cmask(struct radv_image *image) { if (image->planes[0].surface.bpe > 8 && image->info.samples == 1) { /* Do not enable CMASK for non-MSAA images (fast color clear) * because 128 bit formats are not supported, but FMASK might * still be used. */ return false; } return radv_image_can_enable_dcc_or_cmask(image) && image->info.levels == 1 && image->info.depth == 1 && !image->planes[0].surface.is_linear; } static inline bool radv_image_can_enable_fmask(struct radv_image *image) { return image->info.samples > 1 && vk_format_is_color(image->vk_format); } static inline bool radv_image_can_enable_htile(struct radv_image *image) { return radv_image_has_htile(image) && image->info.levels == 1 && image->info.width * image->info.height >= 8 * 8; } static void radv_image_disable_dcc(struct radv_image *image) { for (unsigned i = 0; i < image->plane_count; ++i) image->planes[i].surface.dcc_size = 0; } static void radv_image_disable_htile(struct radv_image *image) { for (unsigned i = 0; i < image->plane_count; ++i) image->planes[i].surface.htile_size = 0; } VkResult radv_image_create(VkDevice _device, const struct radv_image_create_info *create_info, const VkAllocationCallbacks* alloc, VkImage *pImage) { RADV_FROM_HANDLE(radv_device, device, _device); const VkImageCreateInfo *pCreateInfo = create_info->vk_info; struct radv_image *image = NULL; assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO); const unsigned plane_count = vk_format_get_plane_count(pCreateInfo->format); const size_t image_struct_size = sizeof(*image) + sizeof(struct radv_image_plane) * plane_count; radv_assert(pCreateInfo->mipLevels > 0); radv_assert(pCreateInfo->arrayLayers > 0); radv_assert(pCreateInfo->samples > 0); radv_assert(pCreateInfo->extent.width > 0); radv_assert(pCreateInfo->extent.height > 0); radv_assert(pCreateInfo->extent.depth > 0); image = vk_zalloc2(&device->alloc, alloc, image_struct_size, 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (!image) return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY); image->type = pCreateInfo->imageType; image->info.width = pCreateInfo->extent.width; image->info.height = pCreateInfo->extent.height; image->info.depth = pCreateInfo->extent.depth; image->info.samples = pCreateInfo->samples; image->info.storage_samples = pCreateInfo->samples; image->info.array_size = pCreateInfo->arrayLayers; image->info.levels = pCreateInfo->mipLevels; image->info.num_channels = vk_format_get_nr_components(pCreateInfo->format); image->vk_format = pCreateInfo->format; image->tiling = pCreateInfo->tiling; image->usage = pCreateInfo->usage; image->flags = pCreateInfo->flags; image->exclusive = pCreateInfo->sharingMode == VK_SHARING_MODE_EXCLUSIVE; if (pCreateInfo->sharingMode == VK_SHARING_MODE_CONCURRENT) { for (uint32_t i = 0; i < pCreateInfo->queueFamilyIndexCount; ++i) if (pCreateInfo->pQueueFamilyIndices[i] == VK_QUEUE_FAMILY_EXTERNAL || pCreateInfo->pQueueFamilyIndices[i] == VK_QUEUE_FAMILY_FOREIGN_EXT) image->queue_family_mask |= (1u << RADV_MAX_QUEUE_FAMILIES) - 1u; else image->queue_family_mask |= 1u << pCreateInfo->pQueueFamilyIndices[i]; } image->shareable = vk_find_struct_const(pCreateInfo->pNext, EXTERNAL_MEMORY_IMAGE_CREATE_INFO) != NULL; if (!vk_format_is_depth_or_stencil(pCreateInfo->format) && !radv_surface_has_scanout(device, create_info) && !image->shareable) { image->info.surf_index = &device->image_mrt_offset_counter; } image->plane_count = plane_count; image->size = 0; image->alignment = 1; for (unsigned plane = 0; plane < plane_count; ++plane) { struct ac_surf_info info = image->info; radv_init_surface(device, image, &image->planes[plane].surface, plane, create_info); if (plane) { const struct vk_format_description *desc = vk_format_description(pCreateInfo->format); assert(info.width % desc->width_divisor == 0); assert(info.height % desc->height_divisor == 0); info.width /= desc->width_divisor; info.height /= desc->height_divisor; } device->ws->surface_init(device->ws, &info, &image->planes[plane].surface); image->planes[plane].offset = align(image->size, image->planes[plane].surface.surf_alignment); image->size = image->planes[plane].offset + image->planes[plane].surface.surf_size; image->alignment = image->planes[plane].surface.surf_alignment; image->planes[plane].format = vk_format_get_plane_format(image->vk_format, plane); } if (!create_info->no_metadata_planes) { /* Try to enable DCC first. */ if (radv_image_can_enable_dcc(device, image)) { radv_image_alloc_dcc(image); if (image->info.samples > 1) { /* CMASK should be enabled because DCC fast * clear with MSAA needs it. */ assert(radv_image_can_enable_cmask(image)); radv_image_alloc_cmask(device, image); } } else { /* When DCC cannot be enabled, try CMASK. */ radv_image_disable_dcc(image); if (radv_image_can_enable_cmask(image)) { radv_image_alloc_cmask(device, image); } } /* Try to enable FMASK for multisampled images. */ if (radv_image_can_enable_fmask(image)) { radv_image_alloc_fmask(device, image); if (radv_use_tc_compat_cmask_for_image(device, image)) image->tc_compatible_cmask = true; } else { /* Otherwise, try to enable HTILE for depth surfaces. */ if (radv_image_can_enable_htile(image) && !(device->instance->debug_flags & RADV_DEBUG_NO_HIZ)) { image->tc_compatible_htile = image->planes[0].surface.flags & RADEON_SURF_TC_COMPATIBLE_HTILE; radv_image_alloc_htile(device, image); } else { radv_image_disable_htile(image); } } } else { radv_image_disable_dcc(image); radv_image_disable_htile(image); } if (pCreateInfo->flags & VK_IMAGE_CREATE_SPARSE_BINDING_BIT) { image->alignment = MAX2(image->alignment, 4096); image->size = align64(image->size, image->alignment); image->offset = 0; image->bo = device->ws->buffer_create(device->ws, image->size, image->alignment, 0, RADEON_FLAG_VIRTUAL, RADV_BO_PRIORITY_VIRTUAL); if (!image->bo) { vk_free2(&device->alloc, alloc, image); return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY); } } *pImage = radv_image_to_handle(image); return VK_SUCCESS; } static void radv_image_view_make_descriptor(struct radv_image_view *iview, struct radv_device *device, VkFormat vk_format, const VkComponentMapping *components, bool is_storage_image, bool disable_compression, unsigned plane_id, unsigned descriptor_plane_id) { struct radv_image *image = iview->image; struct radv_image_plane *plane = &image->planes[plane_id]; const struct vk_format_description *format_desc = vk_format_description(image->vk_format); bool is_stencil = iview->aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT; uint32_t blk_w; union radv_descriptor *descriptor; uint32_t hw_level = 0; if (is_storage_image) { descriptor = &iview->storage_descriptor; } else { descriptor = &iview->descriptor; } assert(vk_format_get_plane_count(vk_format) == 1); assert(plane->surface.blk_w % vk_format_get_blockwidth(plane->format) == 0); blk_w = plane->surface.blk_w / vk_format_get_blockwidth(plane->format) * vk_format_get_blockwidth(vk_format); if (device->physical_device->rad_info.chip_class >= GFX9) hw_level = iview->base_mip; radv_make_texture_descriptor(device, image, is_storage_image, iview->type, vk_format, components, hw_level, hw_level + iview->level_count - 1, iview->base_layer, iview->base_layer + iview->layer_count - 1, iview->extent.width / (plane_id ? format_desc->width_divisor : 1), iview->extent.height / (plane_id ? format_desc->height_divisor : 1), iview->extent.depth, descriptor->plane_descriptors[descriptor_plane_id], descriptor_plane_id ? NULL : descriptor->fmask_descriptor); const struct legacy_surf_level *base_level_info = NULL; if (device->physical_device->rad_info.chip_class <= GFX9) { if (is_stencil) base_level_info = &plane->surface.u.legacy.stencil_level[iview->base_mip]; else base_level_info = &plane->surface.u.legacy.level[iview->base_mip]; } if (is_storage_image && radv_image_has_dcc(iview->image) && !radv_support_storage_dcc(device->physical_device)) disable_compression = true; si_set_mutable_tex_desc_fields(device, image, base_level_info, plane_id, iview->base_mip, iview->base_mip, blk_w, is_stencil, is_storage_image, disable_compression, descriptor->plane_descriptors[descriptor_plane_id]); } static unsigned radv_plane_from_aspect(VkImageAspectFlags mask) { switch(mask) { case VK_IMAGE_ASPECT_PLANE_1_BIT: return 1; case VK_IMAGE_ASPECT_PLANE_2_BIT: return 2; default: return 0; } } VkFormat radv_get_aspect_format(struct radv_image *image, VkImageAspectFlags mask) { switch(mask) { case VK_IMAGE_ASPECT_PLANE_0_BIT: return image->planes[0].format; case VK_IMAGE_ASPECT_PLANE_1_BIT: return image->planes[1].format; case VK_IMAGE_ASPECT_PLANE_2_BIT: return image->planes[2].format; case VK_IMAGE_ASPECT_STENCIL_BIT: return vk_format_stencil_only(image->vk_format); case VK_IMAGE_ASPECT_DEPTH_BIT: return vk_format_depth_only(image->vk_format); case VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT: return vk_format_depth_only(image->vk_format); default: return image->vk_format; } } void radv_image_view_init(struct radv_image_view *iview, struct radv_device *device, const VkImageViewCreateInfo* pCreateInfo, const struct radv_image_view_extra_create_info* extra_create_info) { RADV_FROM_HANDLE(radv_image, image, pCreateInfo->image); const VkImageSubresourceRange *range = &pCreateInfo->subresourceRange; switch (image->type) { case VK_IMAGE_TYPE_1D: case VK_IMAGE_TYPE_2D: assert(range->baseArrayLayer + radv_get_layerCount(image, range) - 1 <= image->info.array_size); break; case VK_IMAGE_TYPE_3D: assert(range->baseArrayLayer + radv_get_layerCount(image, range) - 1 <= radv_minify(image->info.depth, range->baseMipLevel)); break; default: unreachable("bad VkImageType"); } iview->image = image; iview->bo = image->bo; iview->type = pCreateInfo->viewType; iview->plane_id = radv_plane_from_aspect(pCreateInfo->subresourceRange.aspectMask); iview->aspect_mask = pCreateInfo->subresourceRange.aspectMask; iview->multiple_planes = vk_format_get_plane_count(image->vk_format) > 1 && iview->aspect_mask == VK_IMAGE_ASPECT_COLOR_BIT; iview->vk_format = pCreateInfo->format; if (iview->aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT) { iview->vk_format = vk_format_stencil_only(iview->vk_format); } else if (iview->aspect_mask == VK_IMAGE_ASPECT_DEPTH_BIT) { iview->vk_format = vk_format_depth_only(iview->vk_format); } if (device->physical_device->rad_info.chip_class >= GFX9) { iview->extent = (VkExtent3D) { .width = image->info.width, .height = image->info.height, .depth = image->info.depth, }; } else { iview->extent = (VkExtent3D) { .width = radv_minify(image->info.width , range->baseMipLevel), .height = radv_minify(image->info.height, range->baseMipLevel), .depth = radv_minify(image->info.depth , range->baseMipLevel), }; } if (iview->vk_format != image->planes[iview->plane_id].format) { unsigned view_bw = vk_format_get_blockwidth(iview->vk_format); unsigned view_bh = vk_format_get_blockheight(iview->vk_format); unsigned img_bw = vk_format_get_blockwidth(image->vk_format); unsigned img_bh = vk_format_get_blockheight(image->vk_format); iview->extent.width = round_up_u32(iview->extent.width * view_bw, img_bw); iview->extent.height = round_up_u32(iview->extent.height * view_bh, img_bh); /* Comment ported from amdvlk - * If we have the following image: * Uncompressed pixels Compressed block sizes (4x4) * mip0: 22 x 22 6 x 6 * mip1: 11 x 11 3 x 3 * mip2: 5 x 5 2 x 2 * mip3: 2 x 2 1 x 1 * mip4: 1 x 1 1 x 1 * * On GFX9 the descriptor is always programmed with the WIDTH and HEIGHT of the base level and the HW is * calculating the degradation of the block sizes down the mip-chain as follows (straight-up * divide-by-two integer math): * mip0: 6x6 * mip1: 3x3 * mip2: 1x1 * mip3: 1x1 * * This means that mip2 will be missing texels. * * Fix this by calculating the base mip's width and height, then convert that, and round it * back up to get the level 0 size. * Clamp the converted size between the original values, and next power of two, which * means we don't oversize the image. */ if (device->physical_device->rad_info.chip_class >= GFX9 && vk_format_is_compressed(image->vk_format) && !vk_format_is_compressed(iview->vk_format)) { unsigned lvl_width = radv_minify(image->info.width , range->baseMipLevel); unsigned lvl_height = radv_minify(image->info.height, range->baseMipLevel); lvl_width = round_up_u32(lvl_width * view_bw, img_bw); lvl_height = round_up_u32(lvl_height * view_bh, img_bh); lvl_width <<= range->baseMipLevel; lvl_height <<= range->baseMipLevel; iview->extent.width = CLAMP(lvl_width, iview->extent.width, iview->image->planes[0].surface.u.gfx9.surf_pitch); iview->extent.height = CLAMP(lvl_height, iview->extent.height, iview->image->planes[0].surface.u.gfx9.surf_height); } } iview->base_layer = range->baseArrayLayer; iview->layer_count = radv_get_layerCount(image, range); iview->base_mip = range->baseMipLevel; iview->level_count = radv_get_levelCount(image, range); bool disable_compression = extra_create_info ? extra_create_info->disable_compression: false; for (unsigned i = 0; i < (iview->multiple_planes ? vk_format_get_plane_count(image->vk_format) : 1); ++i) { VkFormat format = vk_format_get_plane_format(iview->vk_format, i); radv_image_view_make_descriptor(iview, device, format, &pCreateInfo->components, false, disable_compression, iview->plane_id + i, i); radv_image_view_make_descriptor(iview, device, format, &pCreateInfo->components, true, disable_compression, iview->plane_id + i, i); } } bool radv_layout_has_htile(const struct radv_image *image, VkImageLayout layout, bool in_render_loop, unsigned queue_mask) { if (radv_image_is_tc_compat_htile(image)) return !in_render_loop; return radv_image_has_htile(image) && (layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL || (layout == VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL && queue_mask == (1u << RADV_QUEUE_GENERAL))); } bool radv_layout_is_htile_compressed(const struct radv_image *image, VkImageLayout layout, bool in_render_loop, unsigned queue_mask) { if (radv_image_is_tc_compat_htile(image)) return !in_render_loop; return radv_image_has_htile(image) && (layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL || (layout == VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL && queue_mask == (1u << RADV_QUEUE_GENERAL))); } bool radv_layout_can_fast_clear(const struct radv_image *image, VkImageLayout layout, bool in_render_loop, unsigned queue_mask) { return layout == VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; } bool radv_layout_dcc_compressed(const struct radv_device *device, const struct radv_image *image, VkImageLayout layout, bool in_render_loop, unsigned queue_mask) { /* Don't compress compute transfer dst, as image stores are not supported. */ if (layout == VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL && (queue_mask & (1u << RADV_QUEUE_COMPUTE)) && !radv_support_storage_dcc(device->physical_device)) return false; return radv_image_has_dcc(image) &&!in_render_loop; } unsigned radv_image_queue_family_mask(const struct radv_image *image, uint32_t family, uint32_t queue_family) { if (!image->exclusive) return image->queue_family_mask; if (family == VK_QUEUE_FAMILY_EXTERNAL || family == VK_QUEUE_FAMILY_FOREIGN_EXT) return (1u << RADV_MAX_QUEUE_FAMILIES) - 1u; if (family == VK_QUEUE_FAMILY_IGNORED) return 1u << queue_family; return 1u << family; } VkResult radv_CreateImage(VkDevice device, const VkImageCreateInfo *pCreateInfo, const VkAllocationCallbacks *pAllocator, VkImage *pImage) { #ifdef ANDROID const VkNativeBufferANDROID *gralloc_info = vk_find_struct_const(pCreateInfo->pNext, NATIVE_BUFFER_ANDROID); if (gralloc_info) return radv_image_from_gralloc(device, pCreateInfo, gralloc_info, pAllocator, pImage); #endif const struct wsi_image_create_info *wsi_info = vk_find_struct_const(pCreateInfo->pNext, WSI_IMAGE_CREATE_INFO_MESA); bool scanout = wsi_info && wsi_info->scanout; return radv_image_create(device, &(struct radv_image_create_info) { .vk_info = pCreateInfo, .scanout = scanout, }, pAllocator, pImage); } void radv_DestroyImage(VkDevice _device, VkImage _image, const VkAllocationCallbacks *pAllocator) { RADV_FROM_HANDLE(radv_device, device, _device); RADV_FROM_HANDLE(radv_image, image, _image); if (!image) return; if (image->flags & VK_IMAGE_CREATE_SPARSE_BINDING_BIT) device->ws->buffer_destroy(image->bo); if (image->owned_memory != VK_NULL_HANDLE) radv_FreeMemory(_device, image->owned_memory, pAllocator); vk_free2(&device->alloc, pAllocator, image); } void radv_GetImageSubresourceLayout( VkDevice _device, VkImage _image, const VkImageSubresource* pSubresource, VkSubresourceLayout* pLayout) { RADV_FROM_HANDLE(radv_image, image, _image); RADV_FROM_HANDLE(radv_device, device, _device); int level = pSubresource->mipLevel; int layer = pSubresource->arrayLayer; unsigned plane_id = radv_plane_from_aspect(pSubresource->aspectMask); struct radv_image_plane *plane = &image->planes[plane_id]; struct radeon_surf *surface = &plane->surface; if (device->physical_device->rad_info.chip_class >= GFX9) { pLayout->offset = plane->offset + surface->u.gfx9.offset[level] + surface->u.gfx9.surf_slice_size * layer; if (image->vk_format == VK_FORMAT_R32G32B32_UINT || image->vk_format == VK_FORMAT_R32G32B32_SINT || image->vk_format == VK_FORMAT_R32G32B32_SFLOAT) { /* Adjust the number of bytes between each row because * the pitch is actually the number of components per * row. */ pLayout->rowPitch = surface->u.gfx9.surf_pitch * surface->bpe / 3; } else { assert(util_is_power_of_two_nonzero(surface->bpe)); pLayout->rowPitch = surface->u.gfx9.surf_pitch * surface->bpe; } pLayout->arrayPitch = surface->u.gfx9.surf_slice_size; pLayout->depthPitch = surface->u.gfx9.surf_slice_size; pLayout->size = surface->u.gfx9.surf_slice_size; if (image->type == VK_IMAGE_TYPE_3D) pLayout->size *= u_minify(image->info.depth, level); } else { pLayout->offset = plane->offset + surface->u.legacy.level[level].offset + (uint64_t)surface->u.legacy.level[level].slice_size_dw * 4 * layer; pLayout->rowPitch = surface->u.legacy.level[level].nblk_x * surface->bpe; pLayout->arrayPitch = (uint64_t)surface->u.legacy.level[level].slice_size_dw * 4; pLayout->depthPitch = (uint64_t)surface->u.legacy.level[level].slice_size_dw * 4; pLayout->size = (uint64_t)surface->u.legacy.level[level].slice_size_dw * 4; if (image->type == VK_IMAGE_TYPE_3D) pLayout->size *= u_minify(image->info.depth, level); } } VkResult radv_CreateImageView(VkDevice _device, const VkImageViewCreateInfo *pCreateInfo, const VkAllocationCallbacks *pAllocator, VkImageView *pView) { RADV_FROM_HANDLE(radv_device, device, _device); struct radv_image_view *view; view = vk_alloc2(&device->alloc, pAllocator, sizeof(*view), 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (view == NULL) return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY); radv_image_view_init(view, device, pCreateInfo, NULL); *pView = radv_image_view_to_handle(view); return VK_SUCCESS; } void radv_DestroyImageView(VkDevice _device, VkImageView _iview, const VkAllocationCallbacks *pAllocator) { RADV_FROM_HANDLE(radv_device, device, _device); RADV_FROM_HANDLE(radv_image_view, iview, _iview); if (!iview) return; vk_free2(&device->alloc, pAllocator, iview); } void radv_buffer_view_init(struct radv_buffer_view *view, struct radv_device *device, const VkBufferViewCreateInfo* pCreateInfo) { RADV_FROM_HANDLE(radv_buffer, buffer, pCreateInfo->buffer); view->bo = buffer->bo; view->range = pCreateInfo->range == VK_WHOLE_SIZE ? buffer->size - pCreateInfo->offset : pCreateInfo->range; view->vk_format = pCreateInfo->format; radv_make_buffer_descriptor(device, buffer, view->vk_format, pCreateInfo->offset, view->range, view->state); } VkResult radv_CreateBufferView(VkDevice _device, const VkBufferViewCreateInfo *pCreateInfo, const VkAllocationCallbacks *pAllocator, VkBufferView *pView) { RADV_FROM_HANDLE(radv_device, device, _device); struct radv_buffer_view *view; view = vk_alloc2(&device->alloc, pAllocator, sizeof(*view), 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); if (!view) return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY); radv_buffer_view_init(view, device, pCreateInfo); *pView = radv_buffer_view_to_handle(view); return VK_SUCCESS; } void radv_DestroyBufferView(VkDevice _device, VkBufferView bufferView, const VkAllocationCallbacks *pAllocator) { RADV_FROM_HANDLE(radv_device, device, _device); RADV_FROM_HANDLE(radv_buffer_view, view, bufferView); if (!view) return; vk_free2(&device->alloc, pAllocator, view); }